Adaptive Video Anomaly Detection by Attention-Based Relational Knowledge Distillation

Detecting anomaly patterns in videos is a challenging task due to complex scenes, huge diversity of anomalies, and fuzzy nature of the task. With advent of technology, tremendous size of visual data is being generated by video surveillance systems, which makes harder to search, analyze, and detect anomalies on video data by human operators. In this paper, we introduce three relational distillation approaches to handle both robust detection of anomalous events and gradual adaptation to different anomaly patterns in new videos while not forgetting anomaly patterns learned from the previous video data. In order to realize these concepts, we propose a unique attention mechanism with feature and relation based knowledge distillation methods. We adapted our knowledge distillation methods to two state-of-the-art models designed for anomaly detection task. Our extensive experiments on two public datasets show that not only our best version model achieves robust performance with a frame-level AUC of 80.22 on UCF-Crime and video-level AUC of 78.20 on RWF-2000 datasets but also the proposed distillation methods improve the performance while reducing catastrophic forgetting problem.